Anthropic details Claude watermark tech for EU AI Act
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic published a blog post explaining its implementation of text watermarking for Claude, using the SynthID-Text method to embed detectable patterns in AI-generated responses without affecting readability.
- Anthropic stated the watermarking relies on low-stakes word choices—like 'overcast' versus 'grey'—that create a hidden signal detectable only with a decoding key, remaining invisible to readers.
- Anthropic clarified that light editing won't remove the watermark, but a full rewrite replacing every word would, at which point the text is no longer meaningfully AI-generated.
- Anthropic noted watermarking has minimal effect on code, as functional constraints limit word choice, though comments and arbitrary naming in code may carry detectable signals.
- Anthropic said watermarking differs from AI detection tools like Pangram’s, which spot stylistic 'tells,' emphasizing watermarks are intentional tags, not inferred patterns.
- Anthropic confirmed other major AI developers have signed the EU AI Act’s Code of Practice and will implement similar watermarking systems for their models.
Why it matters: Developers and enterprises using Claude must now account for embedded watermarks in AI-generated text, which could affect how content is shared, verified, or monetized—especially under EU regulations requiring transparency, with compliance becoming a functional norm across major AI firms.
Ask SkimNews
